Shallot Dataset

This paper presents an edge-based forecasting framework for soil moisture and pH in shallot cultivation by leveraging a modified Temporal Fusion Transformer (TFT) architecture. Unlike conventional TFT, which employs an LSTM encoder\u2013decoder, the proposed model integrates a Cross-Attention mechanism to enhance parallelism, capture long-range dependencies more effectively, and improve interpretability through attention weights. To address noise and variability in IoT sensor data, a Savitzky\u2013Golay (SG) digital filter is applied as preprocessing. The experimental setup involved real-world data collected over a 74-day growth cycle of 300 shallot plants cultivated in Brebes, Indonesia, under three soil moisture conditions (Ideal, Wet, and Dry). The model was trained and deployed on an NVIDIA Jetson Orin edge device to ensure real-time inference. Evaluations used Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Coverage Probability, and Quantile Loss under Elastic Net and Weight Decay regularizations. Results show that Cross-Attention significantly improves predictive stability, especially for pH under Ideal and Wet conditions, while Static Enrichment enhances inter-sensor consistency. SG filtering reduced noise and error variance, leading to more robust forecasts. Benchmark comparisons with LSTM, Informer, Autoformer, and DeepAR confirmed that the proposed TFT\u2013Cross Attention achieves superior accuracy, robustness, and efficiency. With filtering, the model reached low-latency inference (0.12 ms\/sample, 8,505 samples\/s), while the unfiltered variant yielded slightly higher latency (0.13 ms\/sample) and lower throughput (7,651 samples\/s). Nevertheless, this study remains limited to one soil type and agroclimatic condition. Future work will expand to multiple crop varieties, soil textures, and agroclimates to validate generalizability. These findings highlight the potential of the proposed framework to support sustainable precision agriculture and broader smart farming systems